# Scale Without Size: ## How Linktree is Using AI to Accelerate and Do More With Less ### A few weeks ago, [Farnaz Azmoodeh](https://www.linkedin.com/in/farnazazmoodeh/) saw the flash of a Slack notification on her screen. For months, [Linktree](https://linktr.ee/)’s CTO had been tracking the engineering organization’s AI usage through a dashboard she’d vibe coded to better understand adoption and where to bring in new tools. > “We have this amazing principal engineer who’s been the most prolific in terms of just the pure number of PRs he gets in every week. It’s really remarkable how fast he is able to ship,” Azmoodeh says. But when she got the automated Slack ping to check the dashboard that day, for the first time **she saw a new name on top of the PR leaderboard: Devin**. [The AI agent](https://devin.ai/) — the one Linktree’s engineers had likened to a “bad intern” only months earlier — had outpaced a top human developer, sort of. “Of course, the complexity of the PRs that Devin was contributing to were much lower than what our engineers typically accomplish day to day, it's not a real comparison. At least for now though, AI is handling the most repetitive and low-leverage aspects of software engineering, freeing developers to focus on higher-order problem solving,” Azmoodeh says. > But still, the symbolism mattered. For Azmoodeh, it was an early proof point of a new strategy that Linktree started pursuing at the beginning of the year. Going into annual planning for 2025, Linktree’s executive team found itself at a bit of a crossroads. Linktree had started as a simple link-in-bio tool, but over the years, it has quietly [evolved](https://techcrunch.com/2025/04/23/linktree-rolls-out-a-suite-of-monetization-features-for-creators/) into a far more ambitious platform for creators and businesses to run their whole digital presence, supporting 70 million users. > “We’ve been building out a massive array of functionality. But Linktree is a prosumer product. So things really depend on shipping product at high velocity,” says [**Jiaona Zhang**](https://www.linkedin.com/in/jiaona/), who was Linktree's Chief Product Officer during this transformation. ### More specifically, the Linktree team has been building out a two-sided marketplace. On one side, creators needed to be able to monetize through affiliate commerce (earning commissions from third-parties like Amazon, Walmart, and Target) or selling their own courses and digital downloads directly. On the brand side, Linktree wanted to enable companies like Hulu to pay creators for link placement and also run influencer campaigns at scale, even messaging creators directly within the tool. > “When you're building a platform, you don't have the ability to carve off one little sliver. We couldn’t go an inch deep and a mile wide. We needed to be more like 10-feet deep across all these areas to really pull it off in a compelling, high-craft way,” says Zhang. ### That expansive product surface area initially pointed toward significant resourcing needs, which is why early plans called for sizable headcount growth in 2025. **But rather than default to scale, the team decided to experiment to see if they could stay lean**. “At our sub-200 person size, we were at a sweet spot where we hadn’t yet run into some of the common cultural and organizational challenges that come with a larger org, and picking up new skills was still relatively easier. We bet on being able to drive more impact by leaning into AI more aggressively, instead of paying the price of scaling,” says Azmoodeh. ### Now, more than halfway through the year, the strategy seems to be bearing fruit. Here’s a snapshot: But getting to this point required more than just slowing down hiring and sending out an “Everyone needs to use AI” memo. We sat down with Azmoodeh and Zhang to unpack precisely how Linktree pulled off this change, distilling what other companies can learn from their experiment. ## Try many paths, but then prune Linktree’s AI journey had a more conventional start. “We brought in GitHub Copilot first, about two and a half years ago when I first joined the company,” Azmoodeh says. “I generally am not a fan of having a large number of tools. It creates complexity, it increases your risk exposure. It’s just a bad idea,” she says. “But with AI, we were feeling that it was important to lean into the multitude of tools available, so we brought on Cursor soon after it was released, Replit while it was still in beta, then Devin and many more.” > Adoption of tools started slowly. For example, Cursor usage started out at 10% and stayed there the first few weeks, but Azmoodeh felt the numbers were moving in the right direction. To unlock the next level of adoption, she had a series of 1:1 conversations to find out why folks weren’t taking advantage of the new tooling. > “We realized that folks wanted to try out different tools but felt blocked. So we got much faster at reviewing AI tools with our IT and Security teams, so we could onboard them much more quickly than we normally would,” she says. “ **Our security and IT teams committed to 4-6 hours of review per tool once they get access to all the info they need**.” ### Stay ahead. ### Get Applied Intelligence in your inbox. Sometimes this strategy required going the extra mile. “We tried Bolt, Lovable, Replit, and V0 — Lovable’s probably the current favorite. But for a while, there was a group on the design team who preferred to stay within Figma,” says Zhang. “So I pinged [Yuhki](https://review.firstround.com/lessons-in-product-scaling-and-storytelling-from-figmas-cpo/) and managed to get early access to [Figma Make](/content/ai/figma/index.html) just to try to meet people where they are.” But Azmoodeh always intended to reduce the sheer number of AI tools they use. “About a year ago, it became pretty clear that Cursor was the winner for us. They’re a smaller company, they’re much faster moving. The code completion and ability to hold all of the context of the whole code base has really been a game changer for us,” she says. “And it was clear in our usage numbers as well that people were preferring it to the other tools. So once that was the case, **we retired some of these other tools and encouraged the team to use Cursor as their IDE**.” Linktree is going through the same exercise for automated QA testing right now, testing out several tools but also building their own in-house tool. “Many SaaS offerings are increasingly expensive and often add only a thin layer of specialization on top of existing foundational models. Building internally can be a more frugal and precisely tailored path forward, so that’s something we’re exploring more right now,” says Azmoodeh. ## Set your agents up for success Through this trial and (sometimes expensive) error, Linktree discovered which AI tools excel at specific types of work and which ones fail spectacularly at others — at least initially. While some companies have seen mixed results with certain tools, the Linktree team reports getting a lot of mileage out of Devin. At first, Devin couldn’t handle the right level of complexity to be net-useful. “But there’s been a step-function change with the recent updates,” Azmoodeh says. These days, Linktree’s engineers find that Devin excels at many types of repetitive work. It’s proven particularly effective in the following areas: 1. **Prototyping new features and bug fixes:** Linktree’s engineers often use Devin to spike feature ideas or test bug fixes. The output may be far from ship-ready, but it provides a solid starting point. “Having a rough implementation grounds discussions, surfaces trade-offs, and makes project estimation more accurate,” says Azmoodeh. 2. **Upgrading dependencies:** “We had so many dependencies across our code base that were out of date. Our Node.js version was old, and there are so many problems that come along with that. Upgrades are costly. With all these different repos, you have to go and change any APIs that have changed,” she says. “Devin does an excellent job with it. And it does even better if you do the upgrade manually in a couple of repos and then just use those changes as an example, getting Devin to update it everywhere else.” 3. **Identifying and fixing security vulnerabilities:** With a small security team, Linktree faced an endless stream of vulnerabilities potentially replicating. “During code reviews, it’s also good when you prompt it to look for security vulnerabilities, even getting specific about which types to look for in a code base,” she says. “Once one area is fixed and we have an example to show AI, and then we can rinse and repeat across the other code bases and repos, so that's been super handy.” 4. **Debugging:** “Most of the time you can just ask the agents to reason through how a bug can happen. This is still much faster than reproducing the bug, which needs a lot of setup. I’ve also seen engineers prompting Devin to use Datadog’s MCP to assess the scale of a bug’s impact,” she says. ### **Get ahead of failure modes** When these tools fail, however, there tend to be more systemic issues at play. Many engineers were initially quite skeptical. And not without reason. “Linktree’s code base has been around for over seven years, with a lot of tech debt piling up. That all adds up to create complexity,” Azmoodeh says. “ **Fundamentally, AI learns from patterns. When those patterns are a patchwork built up over the years, results can be unpredictable**.” For example, one engineer turned to Devin for help with event logging, and found the task was basically butchered. Digging in, Azmoodeh diagnosed it as challenges with the specific’s case schema design (or lack thereof). “When I looked under the hood, I realized even a human couldn’t easily make sense of things. You need domain knowledge of what these parameters mean. Inconsistent semantics in our schema meant that AI would simply expand these inconsistencies or add to them,” she says. ### Close the adoption gap Still, even with all the tools available to Linktree’s engineers, actual usage wasn’t immediately off to the races. “I’m still constantly watching the numbers. When I take a step back, the curve looks impressive now, but navigating the early phase was super tricky,” says Azmoodeh. Initially, there was a ton of pushback. “In some ways, this is like any other new technology. Engineers are a skeptical bunch, and people get very hooked on their tool of choice. I’m personally that way as well,” she says. “It was hard for me to get used to new data and analytics tools like Amplitude and Hex which we onboarded recently, but once you are over the hump, they bring so much to the table compared to previous generation options. **Old habits die hard though and helping people get over that hump is your number one job as an eng leader right now.**” ### **Pick your targets carefully** Engineering productivity is tricky to define and very gameable. “The moment you declare, ‘We’re going to be looking at lines of code,’ you’ll immediately see more verbosity, more comments,” she says. “If you say you’re tracking the number of PRs, people will start pushing smaller PRs, which could be a good outcome, but you need to think through all the consequences.” But that doesn’t mean sidestepping metrics altogether. Azmoodeh still swears by the classic Charlie Munger quote: “ **Show me the incentives and I'll show you the output**.” That’s why in the near term at least, she chose the deliberately temporary metric of AI tool adoption itself. > “We felt it was important to put emphasis on how much you’re leaning into AI,” she says. “So we’ve been looking at adoption of the top AI tools like Cursor, Devin and Claude Code, and then within each tool, we also look at adoption of different features.” But Azmoodeh plans to keep drilling down here: - **Usage breakdowns:** “We have plans to go deeper into Devin and Cursor adoption by looking at things like diff acceptance rate, Bugbot usage, whether it was used for codegen or understanding the code, and so forth. Cursor doesn’t offer these deeper breakdowns out of the box so we have to build it ourselves,” she says. - **Bug bounty budget:** “Hopefully we can bring this budget down as we prevent more and more vulnerabilities from getting into production, not just through educating our team on how to uphold security but also by heavily utilizing AI,” she says. “We’ve just started down this path, but the idea is that vulnerability detection will be CI blocking in the future.” But this usage metrics focus has a built-in expiration date. “ **I suspect that within three to six months, we won’t even have to look at it anymore**. The whole idea is to force ourselves to get our hands dirty and learn about the technology’s capabilities. What’s the level of complexity it can handle? Where does it fail? Once that familiarity bar is cleared, all we’ll care about is the impact that we’re driving.” > All these input metrics can cause you to lose the forest for the trees. The ultimate output metric that you're going after is dead simple: What's the impact that you’ve been able to drive with these new tools? ### **Leverage usage patterns to start conversations** With a sprawl of tools and a set of metrics to monitor, Azmoodeh set to work to solve visibility. Some tools provided their own analytics, but for comprehensive tracking, she built her own system. “I built a dashboard for myself where I could monitor our team member’s usage of these tools,” she says. When she spotted engineers who weren’t experimenting yet, she crucially didn't wait for them to come around. “I built this flow where below a certain threshold of usage, I could DM the team on Slack and basically say, ‘Hey, I notice that you aren’t using these tools that much yet, can we chat so I can understand what’s holding things back?’” she says. ### **Create peer pressure through public sharing** Lunch and learns and dedicated Slack channels are becoming fairly standard tactics for pushing for AI adoption internally. (Linktree’s channel is dubbed “AI Wins and Whoopsies” to put failure and success on equal footing.) Azmoodeh and Zhang share two other methods they’ve found effective for building social pressure: - **Overhaul standing meetings into AI demos:** “I shifted the product org’s monthly meetings — which used to be product breakdowns and updates on what everyone was working on — to be entirely AI-focused,” says Zhang. “Everyone would go around the horn and share which tools they’re using and what they’ve built. And there’s this social pressure that quickly develops to be able to say you’re using something other than ChatGPT. You can’t say with a straight face that you’re experimenting and trying things out if that’s all you’re doing.” - **Lead by example:** “It really helps to get a couple of folks who are senior and extremely credible to become promoters, and crucially, to show, not just tell. Jiayao Yu stands out as an example. We previously worked together at Snap and he is now one of Linktree’s engineering directors,” says Azmoodeh. “He started just building things on the weekend. His first demo, which he built before even starting at Linktree, was an AI-enabled Q&A tool that would answer questions about the business or creator that each Linktree represented. Even though he’s an engineering director who’s mostly managing, he’s on the tools himself, which emphasizes how important it is to be hands-on at the company.” ## Bring it beyond the eng org Engineering had proven the model, but that same AI-first approach also began spreading throughout the rest of the organization, with some notable results. ### Start with the obvious wins Customer support typically makes for a good first beachhead for AI tooling, and Linktree was no exception. **AI now resolves 50% of customer support tickets, with plans to hit 75% by year-end**. More importantly, Linktree can handle their 80,000 weekly signups without having to significantly scale support hiring. > “We ran POCs with Fin AI, Decagon, and [Lorikeet](https://www.lorikeetcx.ai/),” says Zhang. Measuring on the quality of FAQ responses, ability to customize and integrate with other tools and cost, Lorikeet was a clear winner. “It won because of the ‘knobs and dials.’ We could work with it and make it our own. Everyone tried to sell flexibility but with global rules. There’s more setup required with Lorikeet but it wins long term.” Content moderation delivered another clear-cut victory. With 70 million profiles to moderate for inappropriate content, Linktree faced a challenge that could have easily required hundreds of human reviewers. > “We were relying on basic linear regression models that hit accuracy in the mid-70%. We knew LLMs could push us into the high 80s or low 90s for precision and recall, but it was pretty cost prohibitive at first,” Azmoodeh says. “For one specific adult content classifier, the cost of automated classification of all our links using state of the art LLMs was $520K last year.” But then smaller, more efficient models emerged. By switching to Gemini 2.0 Flash, Linktree achieved the same performance boost as more expensive models, while the cost dropped 25x to $20K. ### The bottom line Linktree’s 2025 annual planning exercise resulted in a bet on scaling the talent of their current team utilising AI instead of solely relying on scaling headcount. There are a few notable takeaways here. One, the turning point came when the company slowed its hiring pace. “When you create that kind of constraint, you’re pushed to fix the underlying system and improve output, not just add more people to work around inefficiencies,” Azmoodeh says. Two, while the tools exist and the playbook is emerging, adoption won't happen organically. Someone has to drive it. “The tools are there, but tools don’t change habits. You need to be willing to champion an unpopular path for a few months. Track the usage, send the annoying DMs, never stop talking about it in meetings,” she says.